Skip to main content

Open Decision Models: Decider, Clef, Kev, JEV

"System One" decision models don't write text. You send a state (text, JSON, sometimes images) and named, typed questions, and get calibrated answers back in one forward pass. They all use the same question schema as Typesafe Jev:

{
"formal": {"type": "noul", "instructions": "Is this formal wear?"},
"category": {"type": "choice", "instructions": "Garment category?",
"criteria": {"dress": "one-piece gowns", "top": "shirts and blouses"}},
"fit": {"type": "score", "instructions": "How fitted is it?",
"criteria": ["loose", "regular", "slim"]}
}

Answers come back as {"model", "answers", "usage"}: noul gives P(yes), choice gives the pick plus probabilities and confidence, and score gives a probability-weighted level with a legend.

At a glance​

Registry idModelHostingImagesNeeds
pplx-deciderPerplexity Decider v1.2 (27B) / Flash v1.2 (4B)Perplexity Decisions API (first-party)up to 8PERPLEXITY_API_KEY
pplx-decider-localperplexity-ai/pplx-decider-v1.1-27byour SGLang serverup to 880 GB-class GPU
clefCloudflare Clef (27B) / Clef-flash (9B)Workers AI (first-party)up to 4CLOUDFLARE_API_TOKEN, CLOUDFLARE_ACCOUNT_ID
clef-localCloudflare/clefyour SGLang serverup to 480 GB+ GPU (tested on H200)
kev-openrouterKev 4BOpenRouter (third-party; Kev has no first-party API)noOPENROUTER_API_KEY
kev-localKev 0.8B / 4B / 9B / 27Bthe Kev repo's kev.serveno4 GB GPU (0.8B) to 80 GB (27B); MLX on Apple Silicon
autotrust-jev-localAutoTrust JEV-35B / JEV-35B-NVFP4 / JEV-27B-VLyour vLLM serve.sh (no hosted API exists)up to 832 GB (NVFP4) to 80 GB GPU

All four model families publish Apache-2.0 weights. Vendor benchmarks are self-reported.

Perplexity Decider​

Decider replaces Qwen3.8-27B's language head with a 255-way answer-code readout, with non-causal attention in its full-attention layers.

export PERPLEXITY_API_KEY=...
opentryon decide --model pplx-decider --questions questions.json --state "Red satin evening gown, floor length"
opentryon decide --model pplx-decider --decider-model pplx-decider-flash-v1.2 --questions questions.json --image dress.jpg
opentryon decide --model pplx-decider --questions questions.json --state "Is this brand in stock at Nordstrom?" --web-grounding
  • Endpoint: POST https://api.perplexity.ai/v1/decisions. Up to 255 options per choice, 10 levels per score.
  • Images go inside state as data-URL parts (PNG, JPEG, WebP). Remote image URLs are downloaded and inlined, because the API rejects them.
  • --web-grounding lets Decider search the web first; searches are billed separately. --max-staleness sets the oldest cached result it may use, in seconds.

Self-hosted: serve the open weights with a nightly SGLang build, then point OpenTryOn at it.

export PPLX_DECIDER_BASE_URL=http://127.0.0.1:30000   # default
opentryon decide --model pplx-decider-local --questions questions.json --state "..." --image dress.jpg

Cloudflare Clef​

Clef adds a joint schema head on a frozen Qwen backbone. The head routes evidence between the state, the questions and their options and scores the whole schema together.

export CLOUDFLARE_API_TOKEN=...      # a Workers AI API token
export CLOUDFLARE_ACCOUNT_ID=...
opentryon decide --model clef --questions questions.json --state "..." --image look.jpg
opentryon decide --model clef --clef-model clef-flash --questions questions.json --state "..."
clefclef-flash
Size27B (Qwen3.8-27B)9B (Qwen3.5-9B)
Context65,536 tokens24,576 tokens
Input price$0.24 per M tokens$0.038 per M tokens

Up to 64 questions and 4 images (PNG, JPEG, WebP; 4 MiB each) per request. Output tokens are free and Cloudflare lists zero data retention. Prices are Cloudflare's published rates at the time of writing.

Self-hosted: run Cloudflare's SGLang image (lmsysorg/sglang:dev-clef, sglang serve --model-path Cloudflare/clef --port 30000), then use --model clef-local (CLEF_BASE_URL, default http://127.0.0.1:30000).

Kev (Jared Palmer)​

Kev is a LoRA plus a small pointer head on Qwen base models. The pointer head scores each option against the question while the state's computation is shared across questions. Kev has no first-party API.

# Hosted, through OpenRouter's System One endpoint (Kev 4B, 8K context)
export OPENROUTER_API_KEY=...
opentryon decide --model kev-openrouter --questions questions.json --state "Customer was charged twice for order 1182"

# Self-hosted, any size, with the Kev repo's server
git clone https://github.com/jaredpalmer/kev.git && cd kev && uv sync --extra serve
uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8008
opentryon decide --model kev-local --questions questions.json --state "..."

Kev is text-only: pass image descriptions or product data as the state. KEV_API_KEY is only needed if you set one on your server.

AutoTrust JEV​

AutoTrust's JEV builds open System One models around a learned decision block (a LoRA plus decision head) on a frozen Qwen backbone, and it reads images. It is a separate company and model from Typesafe's Jev (--model jev).

AutoTrust runs no hosted API and no third-party host serves JEV, so autotrust-jev-local talks to the server you start yourself:

hf download autotrust/JEV-35B --local-dir JEV-35B     # or autotrust/JEV-35B-NVFP4 for a 32 GB GPU
bash JEV-35B/serve.sh # vLLM + POST /v1/decide on :8000
opentryon decide --model autotrust-jev-local --questions questions.json --state "..." --image look.jpg

JEV's /v1/decide answers one question per request, so OpenTryOn sends one request per question and maps the replies onto the shared answer shape:

  • noul: the server scores false and true; the answer is P(true).
  • choice: criteria keys become the options, and their descriptions are added to the question.
  • score: JEV's native score is a fixed 0–5 scale, so your ordered levels are asked as a choice over the levels, and score is the probability-weighted level index.

Python​

from tryon.api.perplexity import PerplexityDeciderAdapter
from tryon.api.cloudflare import ClefAdapter
from tryon.api.systemone import SystemOneAdapter
from tryon.api.autotrust import AutoTrustJevAdapter

questions = {"formal": {"type": "noul", "instructions": "Is this formal wear?"}}

PerplexityDeciderAdapter().decide(questions, state="Red satin gown", image=["gown.jpg"])
ClefAdapter(model="clef-flash").decide(questions, state="Red satin gown")
SystemOneAdapter(preset="kev-local").decide(questions, state="Red satin gown")
AutoTrustJevAdapter().decide(questions, state="Red satin gown")

SystemOneAdapter works with any server that speaks the /v1/systemone contract. Presets: kev-openrouter, kev-local, clef-local and pplx-decider-local.

Already integrated​

Three other decision models from the same wave were already in the decide service: Typesafe Jev (jev), Liquid AI d1 (d1-3b, d1-omni-600m) and the vLLM Semantic Router D3 family (d3-turbo to d3-edge, hosted and local).

Sources​